OpenAI opens its Decisions API to all developers, charging only for input tokens on GPT-6 Luna
The public beta, announced on October 6, runs on GPT-6 Luna and accepts text and images. It returns three structured output types: predicates that estimate the probability a statement is true, choices from a fixed list with confidence scores, and numeric scores against a defined range. There are no charges for output tokens or for cache reads and writes, and OpenAI says decisions come back up to ten times faster than the same model called through the Responses API. The launch follows DevDay and lands during OpenAI's 28 days of developer quality of life updates. Early forum testers were mixed: one reported that a rival classifier was more accurate on nuanced judgment calls and faster per turn, though roughly equal on simple yes or no checks. For builders, routing, moderation, triage and eval grading are often the highest volume calls in an agent stack. A cheap, calibrated, input-only endpoint could replace small fine-tuned classifiers, but teams should benchmark accuracy on their own data before switching.
